Research On Road Elements Detection Technology Based On Deep Learning

In recent years,the combination of deep learning and practical applications integrated tightly.Among them,the detection of road elements based on deep learning has become one of the research hotspots,but the problems encountered are also more seriously,the models cannot meet the needs of high detection accuracy and real-time detection at the same time.In addition,considering the cost and actual needs,road elements detection models must have high scalability in real scenes,and have to meet the needs of increasing or decreasing the object categories that can be detected at any time.Therefore,in order to achieve higher detection accuracy on the basis of real-time detection,we consider both the sample and the road elements detection model based on deep learning,as follows:1.A traffic image augmentation method based on random image mixing up.This article uses images in real traffic scenes for learning,but the road elements in such images are usually in a complex environment with a lot of noise.For example,images collected under bad weather such as cloudy days will cause the roads in the image due to insufficient light.The feature discrimination of the elements is not enough,and the model cannot be fully learned.Repairing such images is time-consuming and the effect cannot be guaranteed.In addition,the collected traffic images cannot cover all possible situations of road elements,and the robustness of the model cannot be guaranteed.Therefore,the traffic image augmentation process based on the mixed neighborhood distribution is performed on the image in the frequency domain,two to four traffic images are randomly stitched,part of the image information is cropped,and the model is forced to recognize the road elements locally and enhance the model’s localization ability.Then perform perspective transformation on the image in the spatial domain,change the pixel position,convert it into the HSV color space,reduce the impact of shadows and other noises,ensure the diversity of road elements,and improve the robustness and detection accuracy of the model.2.Construction of road elements detection model based on one-stage framework.When detecting road elements in a real environment,it is necessary to provide timely feedback to the objects that appear,and to detect them as accurately as possible.Therefore,it is necessary to improve the inference speed and detection accuracy of the model from many aspects.Therefore,in response to many needs,a very efficient model-CSPEfficient Net was designed.First,Efficient Net is used as the basis of the feature fusion network to ensure accurate feature extraction;then,the cross-stage local network idea is incorporated to improve the operating efficiency of the entire network.Secondly,combining the attention mechanism and feature fusion,from the perspective of optimizing features,the feature is weighted to make the model learn more meaningful features.Later,in view of the fact that traditional classification and regression loss functions cannot cope with the extreme imbalance of categories and are not conducive to the detection of small objects,this paper uses the center point instead of the anchor box to detect the object.Finally,from the predicted Gaussian heat map,the center point deviation The moving distance and the predicted road elements size are considered in three aspects,and the final loss function is given.3.Training and testing using datasets from different perspectives(driving perspective and drone perspective)fully proves the applicability and scalability of CSPEfficient Net;conduct a large number of ablation experiments for each part of the research to verify each component of CSPEfficient Net Effectiveness.The CSPEfficient Net proposed in this paper combines the advantages of CSPNet and Efficient Net.It can be applied to low-configuration GPUs.A large number of experiments have proved that it has better accuracy and inference rate for low-configuration GPUs for real-time object detection tasks.Strong advantage.